What is an AI agent? A guide with examples and how to build one (2026)

An AI agent is software that gets to work on its own: you set a goal, the agent picks the steps, uses your tools and data, and only stops once the task is done. So it’s not a glorified chatbot, and not classic automation with fixed if-then rules either. In this article you’ll find the definition, the difference with chatbots and workflows, what agentic AI means, which agents you can use for marketing today and which of them actually work in Belgium. We write from practice: at ClickForest, AI agents we built ourselves with Claude Code and MCP run every day, from the chatbot and the SEO scan on this site to the agents that fact-checked this article. All figures were checked on 3 October 2026.

What is an AI agent?

An AI agent is software in which a language model determines for itself which steps and tools it uses to reach a goal. The agent works in a loop: look at the situation, choose an action, execute it with a tool, evaluate the result and continue until the task is done. That independence sets an agent apart from an ordinary AI tool that gives one answer per question.

Simon Willison, independent open source developer and creator of Datasette, boiled it down to one sentence in September 2025: “An LLM agent runs tools in a loop to achieve a goal.” That goal matters, he adds. There is a stopping condition, so the loop doesn’t run forever.

Anthropic, the company behind Claude, drew the line with classic automation most sharply in December 2024:

“Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”

— Anthropic, engineering blog “Building Effective AI Agents” (Anthropic, 19 December 2024)

Anthropic has since put a warning above that piece itself: much of the tooling it describes has changed since December 2024. The distinction between workflow and agent still stands.

OpenAI applies the same bar in its build guide from April 2025 and lists three fixed building blocks: a model that reasons and decides, tools the agent uses to actually do something (read files, send emails, update your CRM) and instructions that define how it behaves. Miss one and you’re left with a chat window. ClickForest helps Flemish SMEs in Mechelen and Antwerp with AI agents that run on those three building blocks, connected to the systems the business already uses.

What is the difference between a chatbot, a workflow and an AI agent?

A chatbot answers, a workflow runs a fixed script and an AI agent decides for itself. With a chatbot and a workflow every step is set in advance; an agent chooses its own path per situation and uses tools where needed. That lets an agent complete tasks too variable for classic automation.

Mind you: the terms get abused freely in marketing material. In June 2025 Gartner estimated that of the thousands of vendors calling themselves “agentic”, only around 130 offer genuine agent capabilities. They have a word for it: agent washing. An AI chatbot for customer service can be perfectly valuable without being an agent, by the way. It only becomes misleading when a renamed FAQ bot is sold as an autonomous employee.

What is agentic AI and is it the same as an AI agent?

Agentic AI is the umbrella term for AI systems that complete complex tasks with little or no human supervision. An AI agent is the concrete building block of such a system. In practice both terms are used interchangeably: anyone talking about agentic AI is talking about software that works with AI agents.

In November 2024 Deloitte described agentic AI as software that “can complete complex tasks and meet objectives with little or no human supervision”. Since October 2024 Gartner has expected that by 2028 at least 15% of day-to-day work decisions will be made autonomously by agentic AI, and that 33% of enterprise software will include agentic features by then, up from less than 1% in 2024.

Expectations ran high. OpenAI CEO Sam Altman wrote in his January 2025 blog post Reflections: “We believe that, in 2025, we may see the first AI agents ‘join the workforce’ and materially change the output of companies.” Looking back, most companies are not yet scaling agents. In the survey McKinsey ran in May and June 2026, about two in ten respondents say their organisation is really scaling agents. Andrej Karpathy, who uses the agents himself every day, tempered expectations on the Dwarkesh Podcast:

“We have some very early agents that are extremely impressive and that I use daily—Claude and Codex and so on—but I still feel there’s so much work to be done.”

— Andrej Karpathy (Dwarkesh Podcast, 17 October 2025)

There’s a mirror image too: agents that come to buy from you. That’s where the wind turned this year. On 24 March 2026 OpenAI moved away from its standalone Instant Checkout in ChatGPT, because according to the company it didn’t offer the flexibility it had in mind; purchases now happen on the seller’s site. With Shopify, buying through AI channels such as ChatGPT, Gemini or Copilot is currently only possible for US buyers. For a Belgian webshop, being findable is what counts most. How to approach that is covered in our article on agentic commerce.

What can an SME do with AI agents today?

More than you’d think: answering customer queries, preparing quotes, drafting reports, checking content, following up orders and pulling data from different systems together. The rule of thumb: any task that recurs, runs digitally and has a checkable outcome is a candidate. What an agent can’t do is guess what you mean by ‘good’: you define that upfront.

Some examples from our day-to-day work: the chatbot on this site is built on the Claude ecosystem and answers visitor questions from our own knowledge base, our SEO/GEO scan analyses websites autonomously, our CRM has a self-built MCP connection so agents process quotes and customer data directly inside it, and every blog article passes a verification agent that checks figures and quotes against the source. It goes furthest at Kunstbrein, our own AI news site: every morning the site fetches fifty sources, code weighs which story is worth telling, a language model writes the article and it goes online without anyone pressing a button (see the Kunstbrein case). ClickForest helps Flemish SMEs in Mechelen and Antwerp with agents like these, always starting from one concrete task, from marketing automation to customer-service chatbots.

The figures show the timing is right. According to Statbel, 34.5% of Belgian companies with at least ten employees already used at least one AI technology in 2025, rising to three in four among companies with more than 250 employees. The EU average was 20% (Eurostat). On 30 September 2026 Agoria published a survey of executives at Belgian technology companies, with 307 responses: 37% are actually building AI into their processes, up from 27% a year earlier. Yet 85% still see no effect of AI on revenue, and for only 6% is AI an integral part of how the business runs. A Belgian figure specifically on AI agents doesn’t exist yet; Statbel and Agoria measure AI in the broad sense.

Globally, too, agent deployment is still at the start. In the McKinsey report of August 2026, 40% of respondents from organisations with more than $1 billion in revenue say they are scaling agents, up from 27% a year earlier. Among organisations with less than $1 billion in revenue, which are still far from SMEs in the Belgian sense, it stayed stuck at 22%. The gap between trying and doing is exactly where an SME can build a lead now. Bart Steukers, CEO of Agoria, put it like this when presenting that survey:

“We have the talent, the companies and the research needed to succeed in AI. The question is no longer whether Belgium can play a leading role, but how quickly we want to turn that ambition into reality”

— Bart Steukers, CEO of Agoria (Agoria, 30 September 2026)

Anthropic is aiming at smaller companies too. At the launch of the American Claude for Small Business in May 2026, co-founder and president Daniela Amodei said: “Small businesses make up nearly half the American economy, but they’ve never had the resources of bigger companies. AI is the first technology that can finally close that gap.”

Which AI agents can you use for marketing today?

Almost every major platform now ships a ready-made agent: ChatGPT Work, Claude with Cowork, Microsoft Copilot, HubSpot Breeze and Salesforce Agentforce. They look up information, create documents and campaigns, and follow up on leads or customer queries. Far from all of it works in Belgium: Gemini Spark excludes the European Economic Area, and ChatGPT Pro’s dots excluded it at launch too.

Gartner already saw that shift in the numbers in September 2026: “With GenAI firmly in the Trough of Disillusionment in 2026, enterprises are using these simpler embedded AI features from their incumbent software providers”. So even if you don’t build an agent yourself, you get AI built into software you already pay for, and increasingly in the form of an agent. Here is the line-up as of 3 October 2026:

Availability changes fast, so check the vendor’s help page before you choose. Vendor figures such as the 65% of conversations resolved that HubSpot reports for its Customer Agent are self-reported. Which tools make the difference per task is covered in the best unknown AI tools per task and in our overview of AI tools for marketers; the broader context is in our guide to AI marketing.

How do you build an AI agent yourself with Claude Code and MCP?

You need three building blocks: a strong language model, access to your tools and data, and clear instructions. Claude Code delivers the agent loop ready-made, MCP securely connects your own systems, and Skills give the agent reusable methods. Heavier infrastructure isn’t needed for most SME tasks.

Here’s how the pieces fit together:

• Claude Code is the agent environment: it runs in your terminal, editor or browser, can read and write files, run commands and search the web, and splits big work across parallel sub-agents.
• MCP (Model Context Protocol) is the power strip between agent and systems. On 9 December 2025 Anthropic donated the protocol to the Agentic AI Foundation, a fund under the Linux Foundation that it co-founded with Block and OpenAI. At that point Anthropic counted more than 10,000 active public MCP servers, and ChatGPT, Cursor, Gemini, Microsoft Copilot and VS Code already supported it. One connection, every platform.
• Claude Skills are folders with instructions and scripts the agent picks up itself when a task calls for it: your house style, your quote format, your review checklist. Skills have been an open standard since 18 December 2025, and on 3 October 2026 agentskills.io listed 46 tools that support the format, including Cursor, GitHub Copilot and Gemini CLI.
• The Claude Agent SDK gives developers the same agent loop as a standalone building block in Python or TypeScript, for agents that run entirely inside your own product or backend.

No developer in-house? Then Claude Cowork is the entry point: the same agent engine in a regular interface, and since 16 September 2026 simply part of the same conversation as the chat for Pro and Max. From 6 October 2026 new Cowork tasks on those plans run in the cloud, Anthropic announced. More about the app in Claude Desktop explained.

That non-technical use is no niche. According to figures Anthropic published in July 2026, more than 90% of Cowork usage was not software development. In the same figures, covering 1.2 million sessions from the last two weeks of May, TechCrunch saw business processes (33.4%) and content and copywriting (16.4%) in the lead; software development came to 8.7%. Would you rather have this built than figure it out yourself? Then see how we approach building with AI.

Want to know which task in your business is agent-ready first? ClickForest builds AI agents for SMEs in Flanders, from task analysis to a working agent whose output gets checked. Discover the AI agents service.

Do you need n8n or Make to build an AI agent?

No. n8n and Make are workflow platforms: handy for connecting apps with fixed steps, but the agent intelligence comes from the language model, not the platform. Building directly with Claude Code and MCP means fewer moving parts and more control. Choose no-code mainly if you only want to connect existing apps.

Almost every guide online starts at a workflow canvas, and for classic automation that’s fair: n8n shows every step visibly and deliberately keeps a human in the loop. Still, there’s a layer between you and the model, with its own subscriptions, its own limits and an extra place your data passes through. Since MCP became an open standard, an agent can talk to your systems directly and securely, without a workflow platform in between. Anthropic’s own build advice has pointed the same way since December 2024: “Consistently, the most successful implementations use simple, composable patterns rather than complex frameworks.”

Our line at ClickForest: goal and data first, then the tooling. For most agent tasks Claude and MCP will do. If you can draw every step in advance, a regular workflow is often enough.

What does an AI agent cost for an SME?

The cost of an AI agent depends on three factors: the subscription to the underlying model or platform, the build time for the connections to your systems and the management afterwards. Some vendors now charge per result. Most of the thinking goes into defining the task precisely, because a vague goal makes any agent expensive.

An agent doesn’t need expensive infrastructure. Claude Code and Claude Cowork are included in the paid Claude subscriptions (Pro, Max, Team and Enterprise). MCP is an open standard, and many connections are available for free; the Claude Marketplace also lists commercial connectors. Since April 2026 HubSpot has charged per completed task: $0.50 per resolved conversation for the Customer Agent and $1 per lead the Prospecting Agent recommends for outreach. The usage itself costs money too. About 20% of respondents in McKinsey’s 2026 survey say AI-related operating costs, token costs included, constrained their AI use.

So the real question is what the task costs you today. The vendor cases show what’s possible at the top end. At pharma company Novo Nordisk, the time for clinical study documentation dropped from more than 10 weeks to 10 minutes, according to an Anthropic customer case. According to another Anthropic customer case, Rakuten cut the time to market for new features by 79%, from 24 to 5 days. Treat such figures as an indication: they’re self-reported results from large organisations.

“Claude has helped us cut writing times on CSRs by 90% so we can get documentation directly into human hands for review and approval.”

— Waheed Jowiya, Digitalization Strategy Director at Novo Nordisk (customer case on claude.com, accessed 3 October 2026)

How to budget a project like this as an SME without overreaching is covered in our guide on AI consultancy for SMEs. If you would rather have it built, the overview of Belgian agencies that build AI agents shows which eleven demonstrably offer that service, and which two questions separate a real agent from a chatbot with a new label.

What are the risks and pitfalls of AI agents?

The biggest risks are escalating costs, unclear business value and too little risk management. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027 for those reasons. The remedy: start small with a measurable task, build in control, respect the legal disclosure obligation and choose tools that show what the agent does and why.

Gartner made that prediction in June 2025, and analyst Anushree Verma added a warning about the hype:

“Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied.”

— Anushree Verma, Senior Director Analyst at Gartner (Gartner, 25 June 2025)

Our own checks reject plenty as well. At Kunstbrein, half of the written articles don’t make it: across 16 logged runs, 22 were written and 11 published. The gates in code rejected the rest, most often on the fact check. That is the system working.

Then there’s the law. Since 2 August 2026, Article 50 of the EU AI Act applies: anyone offering a chatbot or AI agent that talks directly to people must make sure those people know they’re dealing with AI, unless that’s already obvious. Fines can reach €15 million or 3% of worldwide annual turnover, with room for proportionality for SMEs. Since late July 2026 the rules for high-risk systems have been postponed to 2 December 2027, and those for AI in products to 2 August 2028. AI literacy does remain an obligation: anyone providing or deploying AI systems takes measures to support the AI literacy of their staff, although the law no longer prescribes a fixed level. More on that in our article on what the EU AI Act asks of your chatbot. ClickForest helps Flemish SMEs in Mechelen and Antwerp with AI training that teaches teams how to deploy agents safely.

The pitfalls at a glance:

• Starting from the technology instead of the task. An agent without a clear goal is an expensive demo.
• Autonomy without control. Let an agent run alongside human approval or hard checks for weeks before it’s allowed to execute on its own. We work that way ourselves: every agent output passes a checkpoint, by a human or by a gate in code.
• Treating GDPR and the AI Act as an afterthought. Decide in advance which data the agent sees, where it’s processed, who approves decisions involving personal data and how you let customers know they’re talking to AI.
• Buying agent washing. Ask vendors what the system decides itself. If the answer is “it follows your steps”, you’re buying a workflow.

How do you get started as an SME this month?

Pick one recurring task with clear input and output, give an agent the access it needs, let it run for a week under human control and measure the difference. If it works, you expand to a next task. If it doesn’t, you’ve spent a few days at most learning what needs to improve.

The plan we follow ourselves:

1. List your recurring tasks and score them on volume, error-proneness and how clear the desired outcome is.
2. Pick the most boring task with the clearest outcome. Not the most spectacular one: the most boring. That’s where the fastest return sits.
3. Give the agent access to exactly the data and tools it needs, no more. MCP makes that access scopable.
4. Run a one-week shadow test: the agent works, a human approves. Measure time, quality and errors against the old way.
5. Scale per task, not per department. Every new task gets the same trial period.

What should you remember about AI agents?

An AI agent isn’t a smarter chatbot but a different category of software: a language model that completes tasks on its own with tools and instructions. The building blocks are open and accessible, and the big platforms are now building agents into their own software. Not everything is available in Belgium. What counts is a well-chosen task with checks built in.

With 34.5%, Belgium ranks fourth in the EU for AI use by companies (Eurostat, 2025), but real agent deployment is still at the start everywhere. Automate one task well now and you build a lead while the competition is still watching demos. Looking for a partner to build such an agent or chatbot for you? The overview of Belgian AI chatbot agencies compares nine players per segment. ClickForest helps Flemish SMEs in Mechelen and Antwerp with AI agents: choosing the first task, building the agent and keeping control tight. Book a no-obligation video call and find out which task in your business qualifies first.